The Research Of Intelligent Chat Robot Based On Depth Learning

With the application of deep learning in natural language understanding,word vector representation,machine translation,emotion analysis and Chinese word segmentation,people began to study the key technology of chatting robot and apply deep learning to chatting robot.In recent years,chatting robots have become a very hot research and development direction of artificial intelligence.Currently,Researchers are working on chatting robots in the field of development and opening,Most of them are improved under the framework of Sequence to Sequence(or Encoder-Decoder)in deep learning technology,This article is also under such technical principles,aiming at some problems appeared in the research field of chatting robot are analyzed,and proposes a new chatting robot model,namely the theme of deep learning model and language model(T-DLL Model)is obtained by combining the intelligent chatting robot model.The topic model designed in this paper is a new neural network theme model which combines the traditional topic model LDA and LSTM model.The design of deep learning language model is based on the bidirectional LSTM(Bi-LSTM)Encoder-Decoder model of Attention.The specific research work is as follows:(1)Aiming at the training problem of word vectors,we use word2 vec training tool to express words through distributed representation,and reduce the amount of computation according to the principle of word embedding mechanism "dimensionality reduction".(2)For long distance dependence,Attention model(attention model)and Bi-LTSM(external memory unit)are used.Attention model solves the problem that the traditional Encoder-Decoder framework has only one intermediate semantic vector,which leads to the loss of information and the redundancy of information.At the same time,LSTM model solves the problem of recurrent neural network(RNN)Gradient Vanish(gradient vanishing)and long-term dependency in model training by learning long term information.Bi-LSTM also solves the problem that the LSTM model loses some of the semantic information without taking into account the information below.(3)In order to solve the problem of omnipotent response and enable the robot to generate continuous and meaningful dialogues,this paper proposes the use of vector space to calculate the word vector generated by the word Embedding and the semantic vector which can be used as the topic information is obtained from the neural network theme model the similarity.Add or replace the current theme information according to the similarity value to obtain new current theme information,and then the theme information is combined with the output contextual semantic information of Encoder as the input of Decoder.The chat model with topic information can understand the contextual information that it can solve the problem of generating meaningless dialogues in multiple rounds of conversation which can lead to the difficulty of dialogue.The corpus obtained by crawling the subtitling database is used to train the model proposed in this paper and the deep learning coding library is used to implement the model.The results show that the model proposed in this paper is better than the traditional language model in continuing meaningful dialogue.